✨ integrate crosspaste-mouse as a plugin for cross-device mouse sharing#4222
Open
guiyanakuang wants to merge 34 commits into
Open
✨ integrate crosspaste-mouse as a plugin for cross-device mouse sharing#4222guiyanakuang wants to merge 34 commits into
guiyanakuang wants to merge 34 commits into
Conversation
- Raise events SharedFlow buffer to 256 and document DROP_OLDEST policy
so callers cannot assume every event is delivered.
- close() now interrupts the blocking readLine() via runInterruptible,
closes stdout, and joins the reader job — so AutoCloseable returns
only after the reader coroutine is done (PipedInputStream.close()
alone does not wake a reader parked in read()).
- send() wraps the write+flush in NonCancellable so a mid-send cancel
cannot leave a half-written JSON line on the daemon's stdin.
- Drop the dead 'if Stopped { }' branch; keep the EOF-only contract as a
loop comment.
- Fix spawn() KDoc to describe what it actually throws (IOException
from ProcessBuilder.start, not IllegalStateException).
- Document constructor stream ownership.
…and paired devices
Introduce desktopMouseModule() that registers MouseLayoutStore (backed by DesktopConfigManager via a new typed updateConfig(updater) overload for Map/complex fields) and MouseDaemonManager (with flows derived from DesktopAppConfig.mouseEnabled/mouseListenPort). Launch MouseDaemonManager.run() from CrossPaste startup in non-headless mode only — the daemon is a GUI-oriented feature and serves no purpose in headless/server environments, so desktopMouseModule() is also only added to the non-headless Koin modules list.
Adds en/zh translations for the mouse sharing settings screen and permission dialog, and replaces hardcoded literals in MouseSettingsScreen and MousePermissionDialog with GlobalCopywriter lookups. The nav list entry in AdvancedSettingsContentView already references mouse_settings / mouse_settings_desc, which are now defined. Other locales (de/es/fa/fr/ja/ko/pt/zh_hant) still need translations; i18n_batch_update.sh expects an input file with per-locale values and was not run here because only en/zh copy is available.
Coroutine cancellation via runInterruptible interrupts readLine(), which throws InterruptedIOException. Without catching it, the exception leaks onto the JVM's uncaught-exception queue and breaks adjacent runTest tests with UncaughtExceptionsBeforeTest — surfaced as a flake when MouseLayoutStoreTest or ScreenArrangementViewModelTest ran after MouseDaemonProcessTest in the same JVM fork.
…ns pattern Drop the superfluous updateConfig(updater) lambda overload on DesktopConfigManager — it duplicated machinery the scalar path already has. mouseLayout now flows through the standard scalar copy(key, value) path: the MouseLayoutStore.Backing adapter JSON-encodes on write and decodes on read (via Position's existing @serializable), mirroring how blacklist, sourceExclusions and useNetworkInterfaces are persisted.
upsert/remove previously did snapshot() -> mutate -> set(), allowing concurrent writers to read the same baseline and overwrite each other on write (lost-update race). Replace Backing.set() with update((Map) -> Map) so the read, mutate and write happen atomically inside the backing. - DesktopAppConfigMouseLayoutBacking wraps the whole update in synchronized(updateLock). - Tests' FakeBacking impls use @synchronized update(). - Add a concurrency regression test: 200 concurrent upserts on distinct keys must all survive (would have silently lost most under the old design).
…-context deadlock
close() used runBlocking { cancelAndJoin() } to wait for the reader
coroutine before returning. If close() is called from a coroutine
context (e.g. the MouseDaemonManager collector on a constrained
dispatcher, or from inside the reader itself), runBlocking inside a
coroutine can deadlock the caller's thread.
Replace with a non-suspending scope.cancel(). runInterruptible already
translates the cancel into Thread.interrupt() on the reader's own
Dispatchers.IO thread — the reader unwinds asynchronously. We lose
the strict 'no worker thread outlives close()' property, but the
reader has no external side effects beyond _events.tryEmit into a
DROP_OLDEST buffer, so an async teardown is safe.
… out-of-sync risk Previous design kept a separate _flow MutableStateFlow inside the backing adapter and updated it after configManager.updateConfig. If updateConfig failed and internally rolled back _config.value, our _flow would still advance to the new value — leaving observers in an inconsistent state relative to the persisted config. Refactor: make Backing.flow() return Flow (not MutableStateFlow), and in the desktop adapter derive it directly from configManager.config via map + distinctUntilChanged. Config is now the single source of truth — rollbacks in updateConfig propagate to observers automatically. Test FakeBacking impls declare override return type as Flow<...> explicitly so the compiler generates the required bridge method at the JVM level (AbstractMethodError would fire otherwise).
… boundary
The stop+start fallback (used when the daemon does not advertise
update_layout) emits a Stopped between the two commands that is
session replacement, not a real shutdown. Downstream observers
(MouseDaemonManager state machine, UI) must not treat it as the
daemon stopping.
Route all handle.events through run()'s single collector, which:
- suppresses Stopped while restartInProgress is set,
- forwards everything else to a private SharedFlow,
- clears restartInProgress once the new session's Initialized or
Ready arrives — any Stopped after that is a real shutdown.
Doing the decision at run()'s entry (not with a downstream filter)
removes the race where a lazy filter could be evaluated after the
flag had already been cleared.
Add regression test covering both the suppressed transient and the
subsequent real Stopped.
…g layout Previously, a corrupt mouseLayout field in appConfig.json would fall back to emptyMap() with no log output — users would lose their arrangement with no diagnostic. Add a KotlinLogging warn with the caught exception and the offending JSON (truncated to 200 chars) so operators can see what went wrong. De-duplicated via lastWarnedJson to avoid spam: decode() runs on every snapshot/update and on every emission of the derived flow, so a persistently-corrupt value would otherwise log the same error dozens of times per session.
…updateClientLayout
run() was an ~82 line combine().collect {} lambda mixing three control
paths (disabled teardown, first-enable spawn, layout-change update).
Extract a local Session bundle (client + two Jobs + last inputs) and
three private methods:
- teardownClient(current): stop + close client, cancel jobs, reset state
- spawnClient(inputs): create client, wire event forwarding, start
- updateClientLayout(current, inputs): diff port/peers, call updateLayout
Plus forwardClientEvents() carved out of the inline event-routing
coroutine. run() itself now reads as a four-way state-machine dispatch.
Behavior preserved — all 5 manager tests still pass unchanged.
Method is a test-only injection hook — make it internal so external consumers of the app module can't reach it while desktopTest (same module) retains access. The 'For tests' KDoc alone did not enforce the restriction; the repo has no @VisibleForTesting convention.
….send() Flag for the next person to read this code: the NonCancellable wrap was chosen over caller-cancellation semantics to protect the daemon's JSON-Lines parser from truncated writes. The consequence is that close() waits on writeLock until any in-flight send finishes. Spell that out in the KDoc so nobody 'simplifies' it away.
…vider Adds a synchronous LocalScreensProvider so the canvas can render the user's monitors even before the mouse daemon has emitted Initialized. ScreenInfo gains @transient `name` and `wallpaperPath` populated on the desktop side (NSScreen + wallpaper PNG copy on macOS, AWT fallback elsewhere) — never crosses the daemon IPC boundary.
…unning
The daemon's plugin mode is supposed to keep stdout pure JSONL, but
log lines (with ANSI escapes) occasionally leak there. Treat any line
not starting with '{' as log spillover — log it on our side, but don't
turn it into IpcEvent.Error and surface a red banner in the UI.
Also map IpcEvent.Initialized / Ready to MouseState.Running so a
freshly-started daemon doesn't sit in Starting forever and any
prior transient warning gets cleared.
Replaces the in-app navigation entry with a standalone Compose Window launched from Extensions. The window is resizable and has a 600x500 px minimum size enforced on the underlying AWT window so it can grow to fit denser layouts without ever shrinking past usable. Visibility is driven by a new mouseSettingsWindowVisible state on DesktopAppWindowManager. Removes the now-orphaned commonMain MouseSettings route, the duplicate entry in AdvancedSettings, and the SettingsGraph composable registration so navigation no longer reaches the in-app screen.
MouseSettingsScreen now fills the whole window so the canvas grows with it (both width and height). The canvas itself adds Modifier .clipToBounds() so screen rectangles, wallpapers, and labels at extreme pan/zoom can no longer bleed onto neighbouring UI. Pan and zoom now respect a world-bounding-box-aware clamp: - Pan keeps at least 80 px of bounds visible against every viewport edge so the layout can never drift entirely off-canvas. - Zoom-out floor is fixed at 50 dp per reference 1080p screen via LocalDensity, independent of window size — maximising the window no longer pushes the floor up to fit-the-bounds. - Zoom-in ceiling stays at 1:1. Viewport dimensions now come from BoxWithConstraints.constraints (px) instead of maxWidth/maxHeight (dp). The dp-as-px mismatch had been silently halving the right/bottom pan reach on HiDPI screens. Screen labels rendered with TextMeasurer: font size scales with the rect's smaller side, text is wrapped/ellipsised to fit inside the rect, and a translucent rounded backdrop keeps it legible over any wallpaper.
…erage Renames the entry label in en/zh from "Mouse" / "鼠标共享" to "Keyboard & Mouse" / "键鼠共享" — the feature shares keyboard and mouse, and the surrounding strings already reflected that. Adds the full mouse_settings key set (10 entries: title, desc, canvas help, 5 state strings, switch label, dialog OK) for de, es, fa, fr, ja, ko, pt, and zh_hant; previously they fell back to en.
guiyanakuang
force-pushed
the
claude/elastic-hypatia-c325a1
branch
from
4f257b9 to
5e5fe2c
Compare
The DAO previously used a Channel<String>(UNLIMITED) to notify
getAllSyncRuntimeInfosFlow subscribers of writes. Channel.receiveAsFlow
has fan-out semantics — each emission is consumed by exactly one
collector — so adding a second subscriber caused them to race for each
event. About half the time the wrong collector won, leaving the
SyncManager-driven UI/sync state stale (e.g. removed devices not
disappearing from the list).
Switch to a MutableSharedFlow<Unit> "table changed" signal: every
subscriber sees every change. Subscribers re-read the full table on
each tick via onSubscription { emit(Unit) }.map { getAllSyncRuntimeInfos() },
which also gives them an initial snapshot without a separate code path.
Add a regression test that pins the multi-subscriber behavior.
MouseDaemonManager was reading the paired-device list directly from SyncRuntimeInfoDao.getAllSyncRuntimeInfosFlow(). That works in isolation, but it bypasses the canonical multi-cast view exposed by SyncManager.realTimeSyncRuntimeInfos and conceptually mouse cares about "business peers managed by SyncManager", not raw table rows. Replace the SyncRuntimeInfoDao constructor dependency with a Flow<List<SyncRuntimeInfo>> and wire it to SyncManager.realTimeSyncRuntimeInfos in DesktopMouseModule. Tests pass the same flow shape, so the fake-DAO mock is no longer needed.
The previous default candidate `$HOME/crosspaste-mouse/target/release/crosspaste-mouse` was a developer-machine convention that breaks on Windows (no `.exe` suffix) and has no defined production location. - Read the dev binary path from a new `mouseBinaryPath` key in `development.properties` (added to the tracked template). - In production, look at `pasteAppExePath / crosspaste-mouse[.exe]` — add `MouseDaemonBinary.binaryName(platform)` for the `.exe` suffix. - Move env-aware candidate assembly into the caller (`DesktopMouseModule`) so `MouseDaemonBinary` stays a generic resolver. - Make the "binary not found" error environment-aware: dev points at `development.properties`, prod surfaces the attempted bundled path and tells the user to reinstall. - Keep `-Dcrosspaste.mouse.binary` and `$CROSSPASTE_MOUSE_BIN` as per-launch / per-shell escape hatches. Closes #4303
This was referenced Apr 26, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Integrates the crosspaste-mouse Rust daemon as a subprocess plugin of crosspaste-desktop. Users can now share one keyboard/mouse across paired devices and adjust the virtual screen layout by dragging in a new Settings → Mouse panel.
crosspaste-mouse plugin, speak its JSON Lines protocol (v2) over stdin/stdout withrunInterruptible+NonCancellablefor clean shutdown / atomic writes.MouseDaemonManagerdrives start /UpdateLayout/ stop based onmouseEnabled+mouseListenPortconfig + paired devices (fromSyncRuntimeInfo) + the layout store.DesktopMouseModuleregisters everything and the manager runs inioCoroutineDispatcherat app launch (non-headless only).Trust model: the desktop tells the daemon which peer addresses are valid via already-paired
SyncRuntimeInforows; no cert fingerprint is sent — crosspaste-desktop's own pairing flow is the trust authority, so the daemon's built-inSkipServerVerificationis acceptable here.The full implementation plan lives in docs/superpowers/plans/2026-04-20-crosspaste-mouse-integration.md.
Known gaps / follow-ups
CROSSPASTE_MOUSE_BINenv var or-Dcrosspaste.mouse.binary=<path>system property. Production bundling (Conveyor + cargo cross-compile) is a separate plan.UpdateLayouthot-swap) is still pending. Once it lands,MouseDaemonClientalready prefers the nativeupdate_layoutcommand when advertised viacapabilities.mouse_settings*keys viai18n_batch_update.sh.Test plan
./gradlew ktlintCheckclean./gradlew app:desktopTest --tests "com.crosspaste.mouse.*" --tests "com.crosspaste.ui.mouse.*"— 33/33 passing locallycd ~/crosspaste-mouse && cargo build --releaseCROSSPASTE_MOUSE_BIN=~/crosspaste-mouse/target/release/crosspaste-mouse ./gradlew app:run